While Others Wait for Xbox Labs to Reinvent the Wheel, I Built an Observation Station with WorkBuddy

I came across the news about Microsoft setting up Xbox Labs at 3 a.m. while revising a design doc. Rumor has it this division doesn't build wheels—it builds a "wheel observatory," taking the wild experimental stuff that's been popping up lately, like AI decompilation and vibe coding, and turning it into something players and developers can actually use. One line from Asha Sharma's internal memo really stuck with me:

The boundary between creating and playing is liquefying.

After reading it, I wasn't anxious, because it feels like I've already run a mini version of this approach. My setup is the WorkBuddy desktop app plus Feishu Sheets, and my workspace also has a Unity-exported UI config table hanging in it. It might not work for everyone, but the approach should be copyable.

The smartest thing about Xbox Labs is that it doesn't produce models—it only domesticates energy. Same on my end. I don't expect WorkBuddy to write mods for me. I just have it organize scattered industry material, reference cases, and config changes into a daily-checkable status table. Turning messy information into traceable status—that's something it does very reliably. I've used WorkBuddy for over two months, and this is still the part that performs best.

About a month ago, I was organizing video storyboards and just told WorkBuddy, "Turn these articles into a table." At first glance the output looked fine, but the fields were different every time—today it's called source, tomorrow it's called origin—and later nothing matched up, so it was a total waste.

Later I learned my lesson: Schema must come first. To explain the term, a Schema is a field contract—you first lock down what columns the table has, what each column is called, and what values are allowed, then let the AI fill it in.

Specific steps: open WorkBuddy, create a new workspace, select "Table" on the left, and manually type in the column headers: Item ID, Title, Source, Keywords, Takeaways, Status. After that, click workspace settings in the top right and set this table as the "standard output template." Once that's done, all future tasks will be filled into this mold.

The Status column is the soul. It only allows three values: To Read, Archived, To Reproduce. It is not allowed to improvise. I learned this the hard way—previously I didn't restrict the value range, and it filled in sentences like "worth paying attention to," which couldn't even be filtered.

After the template is set, click "Automation" on the left and create a scheduled task. I set the trigger time to 7 a.m. every day, a bit earlier than when I wake up, so by the time I sit down at my desk, the table is already filled in.

Don't write long natural-language descriptions in the task. I directly reference the template table's field names and tell it to break things down according to the four columns Title, Source, Keywords, Takeaways. My first failure was because I wrote "topic" instead of "keywords," so that entire column was empty, while the other columns were filled in happily. Field names must match exactly—remember this pitfall.

The expected result is this: the next morning when I open the workspace, there are a dozen more rows in the table, and every row's Status is To Read.

Let me say one separate thing about permissions. I set the workspace to "only I can edit, team members can view." Colleagues can only look, not change, which prevents someone from casually altering the template headers. If you want to divide work, you can give different people edit rights to different folders, but the template table itself should have only one administrator. That's the bottom line.

Daily maintenance isn't complicated either. Every morning I spend ten minutes going through it, changing each To Read to Archived or To Reproduce. Rows that fail will automatically be marked red; just click retry. I configure rules at the global level so I don't have to reset them in every new conversation—they take effect automatically like a passive skill. Two weeks ago I also wrote a post about where rules should actually be written, and the conclusion hasn't changed: global first.

Xbox Labs is betting on turning the impossible into something playable. My bet is smaller—turning a tangled mess into a table that automatically grows every morning. The scales are far apart, but the method is the same: define the contract first, then let the AI fill it in.